Channel-Adaptive Temporal Modeling with Grouped Convolutional Network and Informer for Seizure Detection and Prediction

Abstract

Reliable identification and early prediction of epileptic seizures play a critical role in improving patient outcomes and supporting timely therapeutic decision-making. To address the challenges of spatiotemporal feature coupling and long-range dependency modeling in multi-channel EEG signals, this paper proposes a channel-adaptive temporal modeling framework named GCNN-Informer. The framework follows a three-stage pipeline: a grouped convolutional network (GCNN) extracts local features and selects the five most discriminative channels; a convolutional block attention module (CBAM) performs dual spatiotemporal enhancement; and a lightweight Informer encoder captures global temporal dependencies via the ProbSparse self-attention mechanism. Experiments on the CHB-MIT dataset, conducted under a patient-specific evaluation scheme, demonstrated that the proposed method achieved accuracies of 97.33% for seizure detection and 97.28% for seizure prediction, with a low false detection rate (FDR) of 0.30/h and a false prediction rate (FPR) of 0.03/h. Grad-CAM visualization further shows that as EEG signals evolve from normal/interictal to preictal and then to ictal stages, the model’s attention heatmap progressively transitions from locally sparse highlighting to continuous block-like highlighting, and finally to dense deep-red responses—a pattern that closely aligns with clinical electrophysiological dynamics. This study provides an accurate and generalizable deep–learning solution for real–time epileptic monitoring and early–warning systems.